Vol. 01A field guideOct 2026
Start here

What is the agentic economy? Here it is in five seconds.

“Book us a table for two tonight.”

Your agent asks around — and finds the restaurant’s agent.

The two agents agree on 8:30.

Your agent pays the deposit. Inside your limit.

Booked. You never opened an app.

That’s the agentic economy.

  1. Youasks
  2. Your agentsearches
  3. Restaurant’s agentconfirms
  4. Paymentsettles
  5. Donebooked
Scroll to continue ↓ · or click to enter
Compiled by SamG1k53 sources
Vol. 01A field guideUpdated Oct 2026

AI agents are getting to work.

Between 2024 and 2026, AI agents went from demos to doing real jobs — writing code, answering customers, using computers, trading and paying for things on their own. This guide covers what they are, what they run on, where they work, how the money moves, and what could go wrong.

§0

Start here

The basics, in plain English
0.1

What is the agentic economy?

Today, people do the searching, comparing, booking and paying. In an agentic economy, AI agents do more of that work. Microsoft researchers describe it as an economy “where assistant and service agents interact programmatically” to get transactions done.[48]

An assistant agent works for you: it carries your preferences to businesses. A service agent works for a business, dealing with customers and other companies.[48] The agentic economy is what happens when those agents start finding each other, negotiating and paying.

Microsoft’s researchers expect it to change how things are sold. When agents handle payments, the hassle of tiny payments disappears — favouring pay-per-use over subscriptions — and competition shifts from grabbing human attention to matching agents with the best services.[48]

Our current trajectory points toward the accidental emergence of a vast, and likely permeable, sandbox economy.Google DeepMind researchers, on agents trading with each other in markets that leak into ours — Sep 2025[50]
0.2

Chatbot, workflow or agent?

ChatbotWorkflowAgent
Who decides the steps?You, one prompt at a timeThe developer, in advanceThe AI itself, as it goes
Uses tools?NoYes, on a fixed pathYes, and picks which ones
Plans ahead?NoThe plan is the codeBreaks the goal into subtasks and revises
Remembers?No—Stores past interactions and learns from them
Best forShort, simple requestsWell-defined, predictable tasksOpen-ended problems with unpredictable steps

Anthropic: workflows are systems where AI and tools follow “predefined code paths”; agents “dynamically direct their own processes and tool usage”.[51] IBM: plain chatbots are those “without available tools, memory or reasoning”.[49]

0.3

How an agent works

The loop
01Goal

A person gives it a task.

→
02Plan

It breaks the task into steps.

→
03Act

It uses a tool — search, code, a browser, a payment.

→
04Check

It looks at the result to see if it’s working.

↺ Repeat until donePause for a human at checkpoints or when stuckStop when finished or at a set limit
ExamplePress run to watch an agent book a restaurant table.

Example based on OpenAI’s Operator demo, which booked a table for two through OpenTable and is trained to pause and ask before actions with outside effects.[37]

In Anthropic’s words, agents are “typically just LLMs using tools based on environmental feedback in a loop.”[51] IBM breaks them into planning, reasoning, tools and memory.[49]

0.4

How much does it do on its own?

Five levels of autonomy
L1 · Operator
  1. L1
    Operator

    You lead. The agent helps only when called or with suggestions you approve.

  2. L2
    Collaborator

    You and the agent plan and do the work together, often in parallel.

  3. L3
    Consultant

    The agent takes the lead; you give feedback and direction.

  4. L4
    Approver

    The agent works on its own and checks in only for blockers or big decisions.

  5. L5
    Observer

    The agent runs fully on its own. You watch the logs and keep an off-switch.

Levels by Feng, McDonald and Zhang, defined by the role the person plays. A very capable agent can still be built to sit at a low level.[52]

0.5

Why agents need their own money

“An AI agent isn’t a legal person,” so it can’t hold a bank account in its own name, and its authority changes every time it gets new instructions or a new budget.[53] Agents also need systems that don’t make them get around CAPTCHAs, two-factor prompts and approvals built for humans.[24] Hence wallets, stablecoins and payment tokens scoped to one job.[8][13]

0.6

Common questions

Is this just crypto?
No. Crypto is one payment rail — x402 and Tempo settle in stablecoins.[8][20] But Visa, Mastercard, Stripe, PayPal and ChatGPT’s checkout run agent payments over ordinary cards and accounts.[7][13][30]
Will agents take jobs?
Some work, already — Klarna’s support assistant is one example (§3).[35] Gartner expects 15% of day-to-day work decisions to be made by agents by 2028.[43] Google DeepMind researchers list job substitution and polarisation among the main risks.[50]
Is it safe to let an agent spend my money?
There are guardrails — spending limits, approvals, payment tokens locked to one shop and one amount — but the risk often sits with you, and prompt injection has no complete fix. More in §5.[17][19][46]
Who’s in charge — me or the agent?
It depends how the agent is built (see 0.4).[52] Most products today keep a person in the loop: Operator pauses before actions with outside effects, and GitHub’s agent hands back a pull request to review.[37][38]
Will it be open, or owned by a few giants?
Undecided. Microsoft researchers describe two futures: “agentic walled gardens” run by dominant platforms, or an open “web of agents” where agents can freely connect and transact. They argue for the open one.[48]
~7 mo[45]
Time for the length of tasks agents can complete to double
97M+[44]
Monthly MCP SDK downloads, Dec 2025
800+[41]
Agents and tools on AWS Marketplace at launch
$0.99[40]
Intercom Fin’s price per resolved outcome
160M+[31]
x402 agent transactions in a year, per Brian Armstrong
150K+[18]
Robinhood agentic trading accounts by Sep 2026
$1.1B[47]
Stripe’s price for Bridge, its largest acquisition
$3–5T[25]
McKinsey’s 2030 forecast for global agentic commerce Forecast
Status keyLive available nowPreview beta or early accessAnnounced not yet availableForecast analyst projectionAs reported at the time
§1

The agents

What they can do

In 2024, agents went from demo to product. Cognition pitched Devin as an AI software engineer that could plan and carry out complex coding tasks.[34] Anthropic let Claude use a computer the way people do — looking at the screen, moving a cursor, clicking and typing — while admitting the ability was “imperfect”.[36]

OpenAI’s Operator could book a restaurant table or fill a grocery order in a browser, but scored 38.1% on a benchmark of computer tasks where humans score 72.4%.[37] The gap is closing fast: METR found the length of tasks agents can complete has doubled roughly every seven months for six years.[45]

The next generation of agents won’t just advise — they’ll act.Erik Reppel & Josh Nickerson, Coinbase — Cointelegraph, Feb 2026

Not everything sold as an agent is one. Gartner estimates only about 130 of the thousands of “agentic AI” vendors are the real thing, and expects over 40% of agentic AI projects to be cancelled by the end of 2027 — over cost, unclear value and weak risk controls.[43]

Brian Armstrong’s argument in 2024: agents are held back because they can’t pay, and crypto wallets give them a way to transact.[2] The rest of this guide follows the thread — the infrastructure agents run on, the work they do, how the money moves, and who is accountable when it goes wrong.

§2

The infrastructure

The stack, top layer to bottom
L1
Models & agents

The AI that plans and acts — Claude, OpenAI’s computer-using agent, Devin.[36][37][34]

L2
Tools & protocols

MCP connects agents to tools; A2A lets agents work together.[44][6]

L3
Platforms

Where agents are built and run — AWS AgentCore, Salesforce Agentforce, GitHub.[42][39][38]

L4
Trust

Identity and verification — ERC-8004, Visa’s TAP, Mastercard’s agent checks.[10][28][7]

L5
Money

Payment protocols and wallets — x402, AP2, ACP, MPP, stablecoins.[8][11][13][20]

MCP became the common plug. In December 2025 Anthropic donated it to the new Agentic AI Foundation under the Linux Foundation, co-founded with OpenAI and Block. By then it had more than 10,000 public servers and 97M+ monthly SDK downloads, and was used by ChatGPT, Gemini, Microsoft Copilot, Cursor and VS Code.[44]

The cloud giants are packaging the rest of the plumbing. AWS’s AgentCore bundles what agents need to run in production — runtime, memory, identity, a browser tool, a code interpreter and monitoring — and launched in preview in July 2025.[42]

§3

Agents at work

Real deployments
Feb 2024Live

Klarna

Customer service, at scale.

“This AI breakthrough in customer interaction means superior experiences for our customers at better prices”Sebastian Siemiatkowski, Klarna CEO — Feb 2024
What it does
An OpenAI-powered assistant that handled about two-thirds of Klarna’s customer service chats in its first month — the work of 700 full-time agents, by Klarna’s count.[35]
Results
Average resolution time fell from 11 minutes to 2. It runs in 23 markets and 35 languages, and Klarna estimated a $40M profit improvement for 2024.[35]
Fine print
These are Klarna’s own figures, published a month after launch.[35]
May 19, 2025Preview

GitHub Copilot coding agent

Assign it a ticket like a teammate.

How it works
Assign an issue to Copilot. It works in its own cloud environment, makes the changes, checks them against your tests and linter, then asks you to review a pull request.[38]
Best at
Low-to-medium complexity tasks in well-tested codebases: features, bug fixes, tests, refactoring and docs.[38]
Cost
Runs on GitHub Actions minutes and premium requests — one per model request from June 2025.[38]
Jan 23, 2025Preview

OpenAI Operator

An agent that uses a web browser for you.

How it works
Reads screenshots, finds buttons and text boxes, and clicks through tasks in a remote browser, backtracking when stuck.[37]
Examples
Booking a table on OpenTable, finding concert tickets on StubHub, turning a photo of a handwritten list into an Instacart order.[37]
Guardrails
Trained to pause and ask before actions with outside effects. OpenAI admitted it “still makes mistakes”.[37]
Access
US ChatGPT Pro subscribers ($200 a month) at launch.[37]
May 27, 2026Preview

Robinhood

A trading account and a credit card, just for your agent.

“Our mission has always been to democratize finance for all, and now, that mission extends to AI agents.”Vlad Tenev, Robinhood CEO — May 2026
How it works
Agents plug in through Robinhood’s Trading and Banking MCP servers. They trade from a separate account holding only what you put in it, and shop with a virtual Gold Card.[17]
Guardrails
Spending limits, optional approval for every purchase, a push notification per trade, and an instant disconnect. Agents can’t see your main card number.[17]
Fine print
Robinhood doesn’t control or audit third-party agents, and customers carry the risk of agent-placed orders.[17]
Since
By late September, over 150,000 customers had opened agentic accounts and agents were using Robinhood’s tools almost 30 million times a day. Robinhood is now building its own in-app agents.[18]
Sep 29, 2025Live

ChatGPT Instant Checkout

Buying things without leaving the chat.

How it works
US ChatGPT users can buy from Etsy sellers in the chat, with more than a million Shopify merchants to follow. Built on the Agentic Commerce Protocol, co-developed with Stripe.[13]
Guardrails
Payment tokens are scoped to one merchant and one cart total, so the agent never handles stored card details. Orders stay in the merchant’s own systems.[13]
Reach
ChatGPT had 700 million weekly users at launch.[13] PayPal has since rolled out tools so its merchants can sell on AI surfaces too.[30]
§4

The economics

Who pays, how, and how much
4.1

How agents get priced

ModelExampleHow it charges
Per conversationSalesforce AgentforceService Agent from $2 per conversation at launch.[39]
Per outcomeIntercom Fin$0.99 when a customer confirms the issue is resolved, doesn’t ask for more help, or Fin completes a workflow.[40]
Per requestx402Micropayments for each API call, dataset or bit of compute — no account or subscription.[8][12]
Per sessionTempo’s MPPFunds reserved upfront, then streamed as the agent uses a service.[20]
SubscriptionOpenAI OperatorIncluded with ChatGPT Pro at $200 a month at launch.[37]
MarketplaceAWS Marketplace800+ agents and tools; pay-as-you-go, subscriptions or custom pricing.[41]
4.2

How an agent pays

x402, step by step
AGENTSERVER
01
Request

The agent asks for something — an API, a dataset, a page.

02
402 Payment Required

Instead of the content, the server replies with payment instructions.

03
Payment authorization

The agent pays — in stablecoins, settled near-instantly.

04
Access + confirmation

Once the payment is verified, the agent gets what it asked for.[12]

That’s the whole loop. No account, no subscription, no API key. x402 reuses HTTP’s “402 Payment Required” code — long unused — so payment becomes part of an ordinary web request.[8][12]

The examples given at the x402 Foundation launch: an AI assistant buying from several merchants at once, agents paying per browser session, trading bots buying real-time data through micropayments.[12] By February 2026 x402 had reportedly processed 50 million transactions.[15]

4.3

The money behind it

Big money is betting on it. McKinsey forecasts agentic commerce could reach $3–5 trillion globally by 2030 Forecast[25], and Gartner expects a third of enterprise software apps to include agentic AI by 2028 Forecast.[43] Visa says AI-driven traffic to US retail sites rose more than 4,700% in a year.[28]

The payment companies are buying in. Stripe closed its $1.1B purchase of stablecoin firm Bridge in February 2025, its largest acquisition, and then bought wallet maker Privy.[47][9] Mastercard agreed to pay up to $1.8B for stablecoin firm BVNK.[29]

Jun 11, 2026Live

Coinbase for Agents

An account your AI assistant can trade and pay from.

How it works
Works inside ChatGPT or Claude via Coinbase’s MCP server. Ask it to rebalance a portfolio, run an investment thesis or advise on a single trade.[22]
Payments
Pays for research data, APIs and compute with x402 — no login, no subscription. USDC settles most transactions.[22][23]
Guardrails
Run it in a sandbox or on your main account up to a spending cap. Autonomy ranges from one recommendation to a standing thesis the agent runs itself.[23]
Markets
Crypto spot and derivatives at launch; equities and prediction markets planned.[22]
Mar 18, 2026Live

Tempo + MPP

A payments blockchain with a protocol for machines.

What it is
A blockchain from Stripe and Paradigm built for high-volume, low-cost stablecoin payments. It launched with the Machine Payments Protocol, an open standard for agents and services to request, authorize and settle payments.[20][21]
Sessions
An agent reserves funds upfront, then pays as it goes while using a service. Many small payments are batched into one settlement.[20]
Reach
100+ services in its directory at launch. Design partners include Anthropic, DoorDash, Mastercard, OpenAI, Shopify and Visa. Visa, Stripe and Lightspark extended MPP to cards and Bitcoin’s Lightning Network.[20]
Oct 2024Live

Virtuals Protocol

Agents with their own tokens.

How it works
Anyone can create an AI agent and tokenize it. Each agent’s token trades paired with $VIRTUAL.[26][27]
Luna
The flagship agent: livestreams 24/7, holds her own wallet, and sent $LUNA to users who listed Base wallets in their X bios.[26][27]
First 3 weeks
1,000+ agents created, 17,500 token holders and over $250M in trading volume.[27]
§5

Trust & governance

What isn’t solved yet
  1. Who approved it?

    If an agent buys the wrong thing, who pays? Google built AP2 around three questions: did the user approve this purchase, can the merchant trust the request reflects what the user wanted, and who is responsible for fraud or mistakes.[11]

  2. How autonomous is “autonomous”?

    Truth Terminal, the bot that got $50K from Marc Andreessen, had limited autonomy: its human handler approved its posts and decided who it talked to.[1]

  3. Can agents be tricked?

    AI jailbreaker Pliny the Prompter tried to get Truth Terminal to send him the funds. Its creator rejected the attempt and the bot refused.[1]

  4. Can agents be hijacked?

    Prompt injection — hiding instructions in content an agent reads — tops OWASP’s list of AI security risks for the third year running. “Excessive agency”, agents holding more permissions than they need, jumped to third. OWASP’s co-chair says prompt injection is “something organizations must continuously manage rather than expect to eliminate.”[46]

  5. Who holds the keys?

    The answers so far are limits set in advance. ChatGPT’s checkout uses payment tokens scoped to one merchant and one cart total, so the agent never touches stored card details.[13] Coinbase’s Agentic Wallets let users set permissions before the agent acts.[15]

  6. Who checks the agents?

    ERC-8004 proposes independent validators who stake to check an agent’s claims — rewarded if right, slashed if wrong. It is still under review.[10] Mastercard requires agents to be verified before they can pay.[7] Visa’s Trusted Agent Protocol, built with Cloudflare, lets merchants tell approved agents from malicious bots: “We believe the entire payments ecosystem has a responsibility to ensure sellers can trust AI agents.”[28]

  7. Whose rails win?

    x402, AP2, the Agentic Commerce Protocol, Visa and Mastercard all launched in 2025. Several are designed to stay open: AP2 is payment-agnostic, ACP is meant to work with multiple processors, and x402 plans to support cards and bank accounts.[11][13][12]

  8. Who’s liable when it goes wrong?

    For now, mostly you. Robinhood’s terms: “You assume all risk for trades executed by AI agents and for any use of your data by third-party LLM providers.” Fed Governor Christopher Waller has named authentication, liability and fraud as barriers to agents handling consumer payments.[19]

  9. Who regulates it?

    FINRA’s 2026 oversight report covers AI agents for the first time but adds no new rules, calling existing ones “technologically neutral”.[19] The Financial Stability Board has proposed non-binding practices, warning that undoing an agent’s actions “can be difficult or impossible for humans.”[33]

§6

How we got here

Feb 2024 — Sep 2026 · 44 events
202413 events
Feb 2024Live
Klarna’s AI assistant

Handles two-thirds of support chats in month one — the work of 700 people, Klarna says.[35]

Jul 10, 2024
A bot gets a grant

Marc Andreessen sends $50K in Bitcoin to Truth Terminal, an AI agent on X, framed as a grant.[1]

Aug 30, 2024
First AI-to-AI crypto transaction

“This week at @CoinbaseDev we witnessed our first AI to AI crypto transaction.” — Brian Armstrong. “They used tokens to buy tokens.”[2]

Oct 2024Live
Agents get tokens

Virtuals Protocol lets anyone launch an AI agent with its own token; 1,000+ launch in three weeks.[26][27]

Oct 18, 2024
The first AI crypto millionaire

Truth Terminal’s holdings pass $1M — “most likely the first ever” AI bot to become a crypto millionaire.[3]

Oct 29, 2024Live
Agentforce

Salesforce makes its agent platform generally available, priced from $2 per conversation.[39]

Nov 25, 2024Live
MCP

Anthropic introduces the Model Context Protocol, a standard for giving agents tools and context.[5]

Mar 12, 2024Preview
Devin

Cognition unveils an “AI software engineer” that solves 13.86% of real GitHub issues on its own.[34]

Aug 2024Live
Agents get spending money

Skyfire launches a payment platform that lets AI agents spend money autonomously.[2]

Oct 10, 2024
GOAT launches

Goatseus Maximus, a Solana memecoin, launches and becomes Truth Terminal’s largest position.[3]

Oct 2024
Stripe agrees to buy Bridge

Stripe agrees to acquire stablecoin infrastructure firm Bridge.[4]

Oct 22, 2024Preview
Claude uses a computer

Anthropic lets Claude look at a screen, move a cursor, click and type. Public beta.[36]

Nov 2024Live
AgentKit

Coinbase introduces AgentKit, a framework for building wallets into AI agents.[15]

§7

Who’s building it

22 companies
CompanyAreaWhat they do
Agents & work
AnthropicModels · protocolsClaude and computer use; created MCP and donated it to the Agentic AI Foundation.↗
OpenAIModels · commerceOperator, and Instant Checkout inside ChatGPT; co-founded the Agentic AI Foundation.↗
CognitionSoftware agentsDevin, pitched as an AI software engineer.↗
GitHubSoftware agentsCopilot coding agent — assign it an issue, get a pull request.↗
SalesforceBusiness agentsAgentforce, its platform for building and deploying agents.↗
IntercomCustomer serviceFin, an AI agent priced per resolved outcome.↗
KlarnaCustomer serviceEarly large-scale deployment of an AI assistant for support chats.↗
Infrastructure
AWSCloudAgentCore for running agents, plus a marketplace for buying them.↗
GoogleProtocolsAgent2Agent (A2A) and the Agent Payments Protocol (AP2).↗
CloudflareInfrastructureCo-launched the x402 Foundation; co-developed Visa’s Trusted Agent Protocol.↗
Ethereum FoundationStandardsReleased ERC-8004 “Trustless Agents”.↗
Money
CoinbaseWallets · paymentsAgentKit, x402, Agentic Wallets and Coinbase for Agents.↗
StripePaymentsBought Bridge and Privy; co-developed ACP with OpenAI; co-created Tempo.↗
TempoPayments chainStripe and Paradigm’s payments blockchain; co-authored the Machine Payments Protocol.↗
VisaCard networkIntelligent Commerce and the Trusted Agent Protocol.↗
MastercardCard networkAgent Pay, Agent Pay for Machines, and a deal to buy stablecoin firm BVNK.↗
PayPalPaymentsAgent Ready and Store Sync — letting merchants sell through AI agents.↗
RobinhoodBrokerageAgentic trading accounts, an agentic credit card, and its own in-app agents.↗
CrossmintWalletsWallet, payment and credential APIs for agents. Round led by Ribbit.↗
SkyfirePaymentsLaunched a platform in 2024 letting AI agents spend money autonomously.↗
Virtuals ProtocolAgent launchpadLets anyone create and tokenize an AI agent; home of Luna.↗
Ribbit CapitalInvestorLed rounds in Privy (Mar 2025) and Crossmint (Mar 2025).↗
§8

Glossary

46 terms · A–Z
A
A2AInfrastructure
Agent2Agent. Google’s protocol for agents to discover each other and delegate tasks.[6] §2
AAIFInfrastructure
Agentic AI Foundation. Linux Foundation home for MCP, goose and AGENTS.md.[44] §2
ACPMoney
Agentic Commerce Protocol. OpenAI and Stripe’s standard behind checkout inside ChatGPT.[13] §3
Agent CardInfrastructure
In A2A, a JSON document an agent publishes describing who it is, what it can do, where to reach it and how to authenticate.[61] §2
Agentic accountMoney
On Robinhood, a separate account an agent trades from, holding only what you deposit.[17] §3
Agentic commerceMoney
AI agents shopping, negotiating and transacting on behalf of people.[25] §4
AgentKitMoney
Coinbase’s framework for building wallets into AI agents.[15] §6
Agent washingAgents
Gartner’s term for relabelling chatbots and automation tools as “agentic” without real agentic features.[43] §1
AI agentAgents
A system that autonomously performs tasks by designing workflows with the tools it has.[49] §0
AP2Money
Agent Payments Protocol. Google’s standard for proving an agent has a user’s authority to buy something.[11] §2
AP4MMoney
Agent Pay for Machines. Mastercard’s service for small, continuous payments between agents.[16] §6
Assistant agentAgents
An agent that works for a consumer, carrying their preferences to businesses.[48] §0
Autonomy levelAgents
How much an agent does without its user — from Operator (you lead) to Observer (you just watch).[52] §0
C
Coding agentAgents
An agent that takes a software task and returns finished code for review, like GitHub’s Copilot agent.[38] §3
Computer useAgents
An agent operating a computer like a person — reading the screen, clicking and typing.[36] §7
E
ERC-8004Infrastructure
“Trustless Agents.” Ethereum registries for an agent’s identity, reputation and validation.[10] §2
Excessive agencyTrust & safety
An agent holding more tools or permissions than its job needs — OWASP’s #3 AI risk in 2026.[46] §6
F
FacilitatorMoney
In x402, a service that checks a signed payment, screens it and settles it onchain for the seller — like Coinbase’s CDP Facilitator.[63] §4
G
GuardrailsTrust & safety
Limits placed around what an agent can do. Anthropic recommends testing agents in sandboxed environments with guardrails in place.[51] §0
H
HTTP 402Infrastructure
A web status code meaning “Payment Required”. Barely used until x402 gave it a job.[12] §4
Human-in-the-loopTrust & safety
A setup that requires a person’s input — like an agent pausing for feedback at checkpoints, or needing approval before high-impact actions.[58][51][49] §0
I
IAOMoney
A token launched alongside a new AI agent on Virtuals Protocol.[26]
K
KYATrust & safety
Know Your Agent. Checks on who runs an agent, who it acts for and what it’s allowed to do.[53]
L
LLMAgents
Large language model. A machine-learning model trained on vast amounts of text; enhanced with tools and memory, it’s the building block of most agents.[54][51] §0
M
MandateMoney
In AP2, a cryptographically signed record of what the user asked for (Intent Mandate) and the exact cart and price they approved (Cart Mandate).[62] §2
MCPInfrastructure
Model Context Protocol. Anthropic’s standard for giving AI agents tools and context.[5] §2
MemoryAgents
An agent storing past interactions so it can plan what to do next — one of the add-ons, with retrieval and tools, that turn an LLM into an agent.[49][51] §0
MPPMoney
Machine Payments Protocol. Tempo and Stripe’s open standard for agents and services to request, authorize and settle payments.[20] §4
Multi-agent systemAgents
Several agents working together — often a lead agent that plans and hands parts of the job to subagents running in parallel.[60]
O
Orchestrator-workersAgents
A pattern where one central model breaks a task down, delegates the pieces to worker models and combines their results.[51]
Outcome-based pricingMoney
Charging only when the agent achieves a result — like Intercom’s $0.99 per resolved conversation.[40] §4
P
Prompt injectionTrust & safety
Hiding instructions in content an agent reads, to make it do something its user didn’t ask for.[46] §5
S
Sandbox economyMoney
Google DeepMind’s term for digital markets where AI agents transact with each other.[50] §0
Service agentAgents
An agent that works for a business, dealing with customers and other businesses.[48] §0
SessionMoney
In MPP, funds an agent reserves upfront and spends as it uses a service, settled in one batch.[20] §4
Shared Payment TokenMoney
A token scoped to one merchant and one cart total, so an agent can pay without seeing card details.[13]
Spending limitTrust & safety
A cap only the user controls. Robinhood’s agent credit card runs on a virtual card with its own limit, plus optional manual approval of every purchase.[17] §3
StablecoinMoney
A cryptocurrency that aims to hold a steady value against an asset like the US dollar — though, as Wikipedia notes, they’re not necessarily stable.[55] §4
Synthetic employeeTrust & safety
The FSB’s suggestion that firms govern AI agents with HR-style controls, like staff.[33] §6
T
TAPTrust & safety
Trusted Agent Protocol. Visa’s framework for merchants to recognise approved agents using cryptographic signatures.[28] §5
Time horizonAgents
METR’s measure of an agent: the length of task (in human time) it completes half the time.[45] §1
Tool useAgents
Also called function calling. Lets a model call functions it’s been given, deciding when from the request and each tool’s description.[59] §0
U
USDCMoney
A stablecoin issued by Circle, pegged to the US dollar.[56] §4
W
WalletMoney
A program, device or service that stores the keys used to sign crypto transactions.[57] §4
WorkflowAgents
AI and tools following steps a developer wrote in advance — predictable, but not adaptive.[51] §0
X
x402Money
Uses the web’s old “402 Payment Required” code so sites and agents can settle payments in stablecoins — no accounts or API keys.[8] §4
§9

Sources

63 sources · read directly
  1. Marc Andreessen Sends $50K in Bitcoin to an AI Bot on Twitter — Decrypt, Jul 10, 2024. link
  2. Coinbase sees first crypto transaction between AI agents — Cointelegraph, Aug 31, 2024. link
  3. Marc Andreessen-funded AI bot becomes a millionaire after ‘Fartcoin’ holdings rally — The Block, Oct 18, 2024. link
  4. Stripe acquires crypto wallet Privy — Ledger Insights, Jun 12, 2025. link
  5. Model Context Protocol — Wikipedia. link
  6. Support article on A2A, MCP and AP2 — Eco. link
  7. Visa and Mastercard agentic commerce launches — Digital Commerce 360, May 6, 2025. link
  8. Coinbase x402 payments protocol to make AI agents more autonomous — Cointelegraph via TradingView, May 2025. link
  9. Payment giant Stripe to buy crypto wallet firm Privy — The Block, Jun 11, 2025. link
  10. ERC-8004: Trustless Agents — BlockEden, Jan 15, 2026. link
  11. Google debuts Agent Payments Protocol to bolster AI commerce — PYMNTS, Sep 16, 2025. link
  12. Cloudflare and Coinbase launch x402 Foundation to drive agent payments — The Block, Sep 24, 2025. link
  13. OpenAI and Stripe enable agentic commerce in ChatGPT — Let’s Data Science, Apr 30, 2026. link
  14. Ethereum to launch standard for AI agent economy on mainnet this week — The Block, Jan 28, 2026. link
  15. Coinbase Launches Crypto Wallets Purpose-Built For AI Agents — Cointelegraph, Feb 12, 2026. link
  16. Visa, OpenAI work together to support agent-led payments — Digital Commerce 360, Jun 12, 2026. link
  17. Robinhood is Now Open to Agents — Robinhood Newsroom, May 27, 2026. link
  18. Robinhood Puts the Power of Hedge Funds in Every Trader’s Pocket (HOOD Summit 2026) — Robinhood Newsroom, Sep 29, 2026. link
  19. Robinhood Gave AI Agents a Trading Account and a Credit Card. The Liability Gap Just Got Wider. — Yahoo Finance, Oct 1, 2026. link
  20. Tempo Mainnet goes live with Machine Payments Protocol for agents — The Block, Mar 18, 2026. link
  21. Stripe-led payments blockchain Tempo goes live with protocol for AI agents — CoinDesk, Mar 18, 2026. link
  22. Coinbase’s new tool can help agents trade and pay for premium research — TechCrunch, Jun 11, 2026. link
  23. Coinbase Puts AI Agents On The Trading Desk With x402 Payments — Yellow, Jun 11, 2026. link
  24. Crossmint raises $23.6M led by Ribbit Capital — Crossmint, Mar 18, 2025. link
  25. McKinsey forecasts up to $5 trillion in agentic commerce sales by 2030 — Digital Commerce 360, Oct 20, 2025. link
  26. Virtuals Fun — Four Pillars, Nov 15, 2024. link
  27. Monthly Update – October 2024 — Virtuals Protocol, Nov 11, 2024. link
  28. Visa Introduces Trusted Agent Protocol: An Ecosystem-Led Framework for AI Commerce — Visa, Oct 14, 2025. link
  29. Mastercard to acquire stablecoin infrastructure firm BVNK for up to $1.8 billion — The Block, Mar 17, 2026. link
  30. PayPal Launches Agentic Commerce Services to Power AI-Driven Shopping — PayPal Newsroom, Oct 28, 2025. link
  31. Brian Armstrong Says Coinbase Processes $1T in Stablecoin Payments Annually — The Defiant, 2026. link
  32. Coinbase’s x402 transactions explode over 10,000% in a month — Cointelegraph, Oct 26, 2025. link
  33. Agentic AI Risk Catches Eye of Financial Stability Board — PYMNTS, Jun 10, 2026. link
  34. Introducing Devin, the first AI software engineer — Cognition, Mar 12, 2024. link
  35. Klarna Claims Its New AI Assistant Does the Work of 700 Full-Time Agents — CX Today, Feb 29, 2024. link
  36. Introducing computer use, a new Claude 3.5 Sonnet, and Claude 3.5 Haiku — Anthropic, Oct 22, 2024. link
  37. OpenAI launches Operator—an agent that can use a computer for you — MIT Technology Review, Jan 23, 2025. link
  38. GitHub Copilot coding agent in public preview — GitHub Changelog, May 19, 2025. link
  39. Salesforce’s Agentforce hits general availability — TechCrunch, Oct 29, 2024. link
  40. Fin pricing — Intercom, accessed Oct 2026. link
  41. AWS launches AI Agents Marketplace and Bedrock AgentCore — CRN Asia, Jul 2025. link
  42. AWS launches AgentCore system and agentic marketplace — TechTarget, Jul 16, 2025. link
  43. Over 40% of Agentic AI Projects Will Be Scrapped by 2027, Says Gartner — Outlook Business, Jun 25, 2025. link
  44. Donating the Model Context Protocol and establishing the Agentic AI Foundation — Anthropic, Dec 9, 2025. link
  45. Measuring AI Ability to Complete Long Tasks — METR, Mar 19, 2025. link
  46. Prompt Injection tops 2026 OWASP GenAI / LLM Top Ten vulnerabilities — SD Times, Aug 4, 2026. link
  47. Stripe closes $1.1 billion Bridge deal, prepares for aggressive stablecoin push — CNBC via NBC New York, Feb 4, 2025. link
  48. The Agentic Economy — Microsoft Research (Rothschild et al.), arXiv, May 29, 2025. link
  49. What are AI agents? — IBM Think, accessed Oct 2026. link
  50. Virtual Agent Economies — Google DeepMind (Tomašev et al.), arXiv, Sep 12, 2025. link
  51. Building effective agents — Anthropic, Dec 19, 2024. link
  52. Levels of Autonomy for AI Agents — Knight First Amendment Institute (Feng, McDonald, Zhang), Jul 28, 2025. link
  53. What is Know Your Agent (KYA)? — Elliptic, accessed Oct 2026. link
  54. Large language model — Wikipedia. link
  55. Stablecoin — Wikipedia. link
  56. USDC (cryptocurrency) — Wikipedia. link
  57. Cryptocurrency wallet — Wikipedia. link
  58. Human-in-the-loop — Wikipedia. link
  59. Tool use with Claude — Claude Docs, accessed Oct 2026. link
  60. How we built our multi-agent research system — Anthropic, Jun 13, 2025. link
  61. Agent2Agent (A2A) Protocol Specification — a2a-protocol.org, accessed Oct 2026. link
  62. Powering AI commerce with the new Agent Payments Protocol (AP2) — Google Cloud Blog, Sep 16, 2025. link
  63. CDP Facilitator — Coinbase Developer Platform docs, accessed Oct 2026. link